Agents stopped clicking the wrong thing
The fast model now asks twice with the list reversed and knows what you asked for. On 104 test screens it clicked the wrong control 0 times, down from 11.
When an agent names a control in its own words, the fast decision model picks it from the screen. We built a test of 421 screens across the web, your Chrome, an agent's Mac, your Mac and your windows, and climbed it one change at a time, keeping only changes that scored better without more wrong clicks.
Three changes won: the model reads the whole screen when the agent's words match no label, it sees what you asked for and the last three steps, and it is asked twice with the list in reverse order, acting only when both answers agree. Wrong clicks on held-out screens went from 10.6% to 0, and refusals of pay, delete and password controls held.